SSFD⁺: A robust two-stage face detector

Face detectors based on deep learning have demonstrated great progress in detecting multi-scale faces by using multi-scale feature maps and input pyramids. However, using input pyramids and multi-scale feature maps increases the training difficulty and complexity of the network. In this paper, we focus on achieving comparable performance and simplifying the network architecture for detecting multi-scale faces. To enable network learning of multi-scale facial features from a single-scale feature map and a single-scale input image: 1) we conducted a comparative study to investigate which layer contributes more to detecting multi-scale faces and 2) we designed and implemented a simple network structure to improve the performance of detecting multi-scale faces by incorporating additional contextual information. SSFD⁺ achieves mAPs of (91.3%, 90.3%, 83.1%) and (92.4%, 90.9%, 83.7%) on the (easy, medium, and hard) subsets of the WIDER FACE validation and testing datasets, respectively, and promising results on the FDDB, PASCAL Faces, and AFW datasets.